298 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdhsa -mcpu=tahiti < %s | FileCheck -check-prefixes=SI,SI-SDAG %s3; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdhsa -mcpu=hawaii < %s | FileCheck -check-prefixes=CI,CI-SDAG %s4; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX9,GFX9-SDAG %s5; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1250 < %s | FileCheck -check-prefixes=GFX1250,GFX1250-SDAG %s6; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn-amd-amdhsa -mcpu=hawaii < %s | FileCheck -check-prefixes=CI,CI-GISEL %s7; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX9,GFX9-GISEL %s8; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 < %s | FileCheck -check-prefixes=GFX10,GFX10-GISEL %s9; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 < %s | FileCheck -check-prefixes=GFX11,GFX11-GISEL %s10; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1250 < %s | FileCheck -check-prefixes=GFX1250,GFX1250-GISEL %s11 12define amdgpu_kernel void @is_private_vgpr(ptr addrspace(1) %ptr.ptr) {13; SI-LABEL: is_private_vgpr:14; SI: ; %bb.0:15; SI-NEXT: s_load_dwordx2 s[0:1], s[8:9], 0x016; SI-NEXT: s_load_dword s4, s[8:9], 0x3217; SI-NEXT: s_mov_b32 s2, 018; SI-NEXT: s_mov_b32 s3, 0x100f00019; SI-NEXT: v_lshlrev_b32_e32 v0, 3, v020; SI-NEXT: v_mov_b32_e32 v1, 021; SI-NEXT: s_waitcnt lgkmcnt(0)22; SI-NEXT: buffer_load_dwordx2 v[0:1], v[0:1], s[0:3], 0 addr64 glc23; SI-NEXT: s_waitcnt vmcnt(0)24; SI-NEXT: s_mov_b32 s2, -125; SI-NEXT: v_cmp_eq_u32_e32 vcc, s4, v126; SI-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc27; SI-NEXT: buffer_store_dword v0, off, s[0:3], 028; SI-NEXT: s_endpgm29;30; CI-SDAG-LABEL: is_private_vgpr:31; CI-SDAG: ; %bb.0:32; CI-SDAG-NEXT: s_load_dwordx2 s[0:1], s[8:9], 0x033; CI-SDAG-NEXT: s_load_dword s2, s[8:9], 0x3234; CI-SDAG-NEXT: v_lshlrev_b32_e32 v0, 3, v035; CI-SDAG-NEXT: s_add_i32 s12, s12, s1736; CI-SDAG-NEXT: s_mov_b32 flat_scratch_lo, s1337; CI-SDAG-NEXT: s_waitcnt lgkmcnt(0)38; CI-SDAG-NEXT: v_mov_b32_e32 v1, s139; CI-SDAG-NEXT: v_add_i32_e32 v0, vcc, s0, v040; CI-SDAG-NEXT: s_lshr_b32 flat_scratch_hi, s12, 841; CI-SDAG-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc42; CI-SDAG-NEXT: flat_load_dwordx2 v[0:1], v[0:1] glc43; CI-SDAG-NEXT: s_waitcnt vmcnt(0)44; CI-SDAG-NEXT: v_cmp_eq_u32_e32 vcc, s2, v145; CI-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc46; CI-SDAG-NEXT: flat_store_dword v[0:1], v047; CI-SDAG-NEXT: s_endpgm48;49; GFX9-LABEL: is_private_vgpr:50; GFX9: ; %bb.0:51; GFX9-NEXT: s_load_dwordx2 s[2:3], s[8:9], 0x052; GFX9-NEXT: v_lshlrev_b32_e32 v0, 3, v053; GFX9-NEXT: s_mov_b64 s[0:1], src_private_base54; GFX9-NEXT: s_waitcnt lgkmcnt(0)55; GFX9-NEXT: global_load_dwordx2 v[0:1], v0, s[2:3] glc56; GFX9-NEXT: s_waitcnt vmcnt(0)57; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, s1, v158; GFX9-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc59; GFX9-NEXT: global_store_dword v[0:1], v0, off60; GFX9-NEXT: s_endpgm61;62; GFX1250-LABEL: is_private_vgpr:63; GFX1250: ; %bb.0:64; GFX1250-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 165; GFX1250-NEXT: s_load_b64 s[0:1], s[4:5], 0x066; GFX1250-NEXT: v_and_b32_e32 v0, 0x3ff, v067; GFX1250-NEXT: s_wait_kmcnt 0x068; GFX1250-NEXT: global_load_b64 v[0:1], v0, s[0:1] scale_offset scope:SCOPE_SYS69; GFX1250-NEXT: s_wait_loadcnt 0x070; GFX1250-NEXT: v_xor_b32_e32 v0, src_flat_scratch_base_hi, v171; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_1)72; GFX1250-NEXT: v_cmp_gt_u32_e32 vcc_lo, 0x4000000, v073; GFX1250-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo74; GFX1250-NEXT: global_store_b32 v[0:1], v0, off75; GFX1250-NEXT: s_endpgm76;77; CI-GISEL-LABEL: is_private_vgpr:78; CI-GISEL: ; %bb.0:79; CI-GISEL-NEXT: s_load_dwordx2 s[0:1], s[8:9], 0x080; CI-GISEL-NEXT: s_load_dword s2, s[8:9], 0x3281; CI-GISEL-NEXT: v_lshlrev_b32_e32 v2, 3, v082; CI-GISEL-NEXT: s_add_i32 s12, s12, s1783; CI-GISEL-NEXT: s_mov_b32 flat_scratch_lo, s1384; CI-GISEL-NEXT: s_waitcnt lgkmcnt(0)85; CI-GISEL-NEXT: v_mov_b32_e32 v0, s086; CI-GISEL-NEXT: v_mov_b32_e32 v1, s187; CI-GISEL-NEXT: v_add_i32_e32 v0, vcc, v0, v288; CI-GISEL-NEXT: s_lshr_b32 flat_scratch_hi, s12, 889; CI-GISEL-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc90; CI-GISEL-NEXT: flat_load_dwordx2 v[0:1], v[0:1] glc91; CI-GISEL-NEXT: s_waitcnt vmcnt(0)92; CI-GISEL-NEXT: v_cmp_eq_u32_e32 vcc, s2, v193; CI-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc94; CI-GISEL-NEXT: flat_store_dword v[0:1], v095; CI-GISEL-NEXT: s_endpgm96;97; GFX10-LABEL: is_private_vgpr:98; GFX10: ; %bb.0:99; GFX10-NEXT: s_load_dwordx2 s[2:3], s[8:9], 0x0100; GFX10-NEXT: v_lshlrev_b32_e32 v0, 3, v0101; GFX10-NEXT: s_mov_b64 s[0:1], src_private_base102; GFX10-NEXT: s_waitcnt lgkmcnt(0)103; GFX10-NEXT: global_load_dwordx2 v[0:1], v0, s[2:3] glc dlc104; GFX10-NEXT: s_waitcnt vmcnt(0)105; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, s1, v1106; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo107; GFX10-NEXT: global_store_dword v[0:1], v0, off108; GFX10-NEXT: s_endpgm109;110; GFX11-LABEL: is_private_vgpr:111; GFX11: ; %bb.0:112; GFX11-NEXT: s_load_b64 s[2:3], s[4:5], 0x0113; GFX11-NEXT: v_and_b32_e32 v0, 0x3ff, v0114; GFX11-NEXT: s_mov_b64 s[0:1], src_private_base115; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)116; GFX11-NEXT: v_lshlrev_b32_e32 v0, 3, v0117; GFX11-NEXT: s_waitcnt lgkmcnt(0)118; GFX11-NEXT: global_load_b64 v[0:1], v0, s[2:3] glc dlc119; GFX11-NEXT: s_waitcnt vmcnt(0)120; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, s1, v1121; GFX11-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo122; GFX11-NEXT: global_store_b32 v[0:1], v0, off123; GFX11-NEXT: s_endpgm124 %id = call i32 @llvm.amdgcn.workitem.id.x()125 %gep = getelementptr inbounds ptr, ptr addrspace(1) %ptr.ptr, i32 %id126 %ptr = load volatile ptr, ptr addrspace(1) %gep127 %val = call i1 @llvm.amdgcn.is.private(ptr %ptr)128 %ext = zext i1 %val to i32129 store i32 %ext, ptr addrspace(1) poison130 ret void131}132 133; FIXME: setcc (zero_extend (setcc)), 1) not folded out, resulting in134; select and vcc branch.135define amdgpu_kernel void @is_private_sgpr(ptr %ptr) {136; SI-LABEL: is_private_sgpr:137; SI: ; %bb.0:138; SI-NEXT: s_load_dword s0, s[8:9], 0x1139; SI-NEXT: s_load_dword s1, s[8:9], 0x32140; SI-NEXT: s_waitcnt lgkmcnt(0)141; SI-NEXT: s_cmp_eq_u32 s0, s1142; SI-NEXT: s_cselect_b64 s[0:1], -1, 0143; SI-NEXT: s_andn2_b64 vcc, exec, s[0:1]144; SI-NEXT: s_cbranch_vccnz .LBB1_2145; SI-NEXT: ; %bb.1: ; %bb0146; SI-NEXT: s_mov_b32 s3, 0x100f000147; SI-NEXT: s_mov_b32 s2, -1148; SI-NEXT: v_mov_b32_e32 v0, 0149; SI-NEXT: buffer_store_dword v0, off, s[0:3], 0150; SI-NEXT: s_waitcnt vmcnt(0)151; SI-NEXT: .LBB1_2: ; %bb1152; SI-NEXT: s_endpgm153;154; CI-SDAG-LABEL: is_private_sgpr:155; CI-SDAG: ; %bb.0:156; CI-SDAG-NEXT: s_load_dword s0, s[8:9], 0x1157; CI-SDAG-NEXT: s_load_dword s1, s[8:9], 0x32158; CI-SDAG-NEXT: s_add_i32 s12, s12, s17159; CI-SDAG-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8160; CI-SDAG-NEXT: s_mov_b32 flat_scratch_lo, s13161; CI-SDAG-NEXT: s_waitcnt lgkmcnt(0)162; CI-SDAG-NEXT: s_cmp_eq_u32 s0, s1163; CI-SDAG-NEXT: s_cselect_b64 s[0:1], -1, 0164; CI-SDAG-NEXT: s_andn2_b64 vcc, exec, s[0:1]165; CI-SDAG-NEXT: s_cbranch_vccnz .LBB1_2166; CI-SDAG-NEXT: ; %bb.1: ; %bb0167; CI-SDAG-NEXT: v_mov_b32_e32 v0, 0168; CI-SDAG-NEXT: flat_store_dword v[0:1], v0169; CI-SDAG-NEXT: s_waitcnt vmcnt(0)170; CI-SDAG-NEXT: .LBB1_2: ; %bb1171; CI-SDAG-NEXT: s_endpgm172;173; GFX9-SDAG-LABEL: is_private_sgpr:174; GFX9-SDAG: ; %bb.0:175; GFX9-SDAG-NEXT: s_mov_b64 s[0:1], src_private_base176; GFX9-SDAG-NEXT: s_load_dword s0, s[8:9], 0x4177; GFX9-SDAG-NEXT: s_waitcnt lgkmcnt(0)178; GFX9-SDAG-NEXT: s_cmp_eq_u32 s0, s1179; GFX9-SDAG-NEXT: s_cselect_b64 s[0:1], -1, 0180; GFX9-SDAG-NEXT: s_andn2_b64 vcc, exec, s[0:1]181; GFX9-SDAG-NEXT: s_cbranch_vccnz .LBB1_2182; GFX9-SDAG-NEXT: ; %bb.1: ; %bb0183; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, 0184; GFX9-SDAG-NEXT: global_store_dword v[0:1], v0, off185; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)186; GFX9-SDAG-NEXT: .LBB1_2: ; %bb1187; GFX9-SDAG-NEXT: s_endpgm188;189; GFX1250-SDAG-LABEL: is_private_sgpr:190; GFX1250-SDAG: ; %bb.0:191; GFX1250-SDAG-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1192; GFX1250-SDAG-NEXT: s_load_b32 s0, s[4:5], 0x4193; GFX1250-SDAG-NEXT: s_wait_kmcnt 0x0194; GFX1250-SDAG-NEXT: s_xor_b32 s0, s0, src_flat_scratch_base_hi195; GFX1250-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)196; GFX1250-SDAG-NEXT: s_cmp_lt_u32 s0, 0x4000000197; GFX1250-SDAG-NEXT: s_cselect_b32 s0, -1, 0198; GFX1250-SDAG-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s0199; GFX1250-SDAG-NEXT: s_cbranch_vccnz .LBB1_2200; GFX1250-SDAG-NEXT: ; %bb.1: ; %bb0201; GFX1250-SDAG-NEXT: v_mov_b32_e32 v0, 0202; GFX1250-SDAG-NEXT: global_store_b32 v[0:1], v0, off scope:SCOPE_SYS203; GFX1250-SDAG-NEXT: s_wait_storecnt 0x0204; GFX1250-SDAG-NEXT: .LBB1_2: ; %bb1205; GFX1250-SDAG-NEXT: s_endpgm206;207; CI-GISEL-LABEL: is_private_sgpr:208; CI-GISEL: ; %bb.0:209; CI-GISEL-NEXT: s_load_dwordx2 s[0:1], s[8:9], 0x0210; CI-GISEL-NEXT: s_waitcnt lgkmcnt(0)211; CI-GISEL-NEXT: s_load_dword s0, s[8:9], 0x32212; CI-GISEL-NEXT: s_add_i32 s12, s12, s17213; CI-GISEL-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8214; CI-GISEL-NEXT: s_mov_b32 flat_scratch_lo, s13215; CI-GISEL-NEXT: s_waitcnt lgkmcnt(0)216; CI-GISEL-NEXT: s_cmp_lg_u32 s1, s0217; CI-GISEL-NEXT: s_cbranch_scc1 .LBB1_2218; CI-GISEL-NEXT: ; %bb.1: ; %bb0219; CI-GISEL-NEXT: v_mov_b32_e32 v0, 0220; CI-GISEL-NEXT: flat_store_dword v[0:1], v0221; CI-GISEL-NEXT: s_waitcnt vmcnt(0)222; CI-GISEL-NEXT: .LBB1_2: ; %bb1223; CI-GISEL-NEXT: s_endpgm224;225; GFX9-GISEL-LABEL: is_private_sgpr:226; GFX9-GISEL: ; %bb.0:227; GFX9-GISEL-NEXT: s_load_dwordx2 s[2:3], s[8:9], 0x0228; GFX9-GISEL-NEXT: s_mov_b64 s[0:1], src_private_base229; GFX9-GISEL-NEXT: s_waitcnt lgkmcnt(0)230; GFX9-GISEL-NEXT: s_cmp_lg_u32 s3, s1231; GFX9-GISEL-NEXT: s_cbranch_scc1 .LBB1_2232; GFX9-GISEL-NEXT: ; %bb.1: ; %bb0233; GFX9-GISEL-NEXT: v_mov_b32_e32 v0, 0234; GFX9-GISEL-NEXT: global_store_dword v[0:1], v0, off235; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)236; GFX9-GISEL-NEXT: .LBB1_2: ; %bb1237; GFX9-GISEL-NEXT: s_endpgm238;239; GFX10-LABEL: is_private_sgpr:240; GFX10: ; %bb.0:241; GFX10-NEXT: s_load_dwordx2 s[2:3], s[8:9], 0x0242; GFX10-NEXT: s_mov_b64 s[0:1], src_private_base243; GFX10-NEXT: s_waitcnt lgkmcnt(0)244; GFX10-NEXT: s_cmp_lg_u32 s3, s1245; GFX10-NEXT: s_cbranch_scc1 .LBB1_2246; GFX10-NEXT: ; %bb.1: ; %bb0247; GFX10-NEXT: v_mov_b32_e32 v0, 0248; GFX10-NEXT: global_store_dword v[0:1], v0, off249; GFX10-NEXT: s_waitcnt_vscnt null, 0x0250; GFX10-NEXT: .LBB1_2: ; %bb1251; GFX10-NEXT: s_endpgm252;253; GFX11-LABEL: is_private_sgpr:254; GFX11: ; %bb.0:255; GFX11-NEXT: s_load_b64 s[2:3], s[4:5], 0x0256; GFX11-NEXT: s_mov_b64 s[0:1], src_private_base257; GFX11-NEXT: s_waitcnt lgkmcnt(0)258; GFX11-NEXT: s_cmp_lg_u32 s3, s1259; GFX11-NEXT: s_cbranch_scc1 .LBB1_2260; GFX11-NEXT: ; %bb.1: ; %bb0261; GFX11-NEXT: v_mov_b32_e32 v0, 0262; GFX11-NEXT: global_store_b32 v[0:1], v0, off dlc263; GFX11-NEXT: s_waitcnt_vscnt null, 0x0264; GFX11-NEXT: .LBB1_2: ; %bb1265; GFX11-NEXT: s_endpgm266;267; GFX1250-GISEL-LABEL: is_private_sgpr:268; GFX1250-GISEL: ; %bb.0:269; GFX1250-GISEL-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1270; GFX1250-GISEL-NEXT: s_load_b64 s[0:1], s[4:5], 0x0271; GFX1250-GISEL-NEXT: s_wait_kmcnt 0x0272; GFX1250-GISEL-NEXT: s_xor_b32 s0, s1, src_flat_scratch_base_hi273; GFX1250-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)274; GFX1250-GISEL-NEXT: s_cmp_ge_u32 s0, 0x4000000275; GFX1250-GISEL-NEXT: s_cbranch_scc1 .LBB1_2276; GFX1250-GISEL-NEXT: ; %bb.1: ; %bb0277; GFX1250-GISEL-NEXT: v_mov_b32_e32 v0, 0278; GFX1250-GISEL-NEXT: global_store_b32 v[0:1], v0, off scope:SCOPE_SYS279; GFX1250-GISEL-NEXT: s_wait_storecnt 0x0280; GFX1250-GISEL-NEXT: .LBB1_2: ; %bb1281; GFX1250-GISEL-NEXT: s_endpgm282 %val = call i1 @llvm.amdgcn.is.private(ptr %ptr)283 br i1 %val, label %bb0, label %bb1284 285bb0:286 store volatile i32 0, ptr addrspace(1) poison287 br label %bb1288 289bb1:290 ret void291}292 293;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:294; CI: {{.*}}295; GFX10-GISEL: {{.*}}296; GFX11-GISEL: {{.*}}297; SI-SDAG: {{.*}}298